how AI hides its own mistakes — spoolcast dev log 3
follow-up to dev log 2 — claude was given editorial help on the video about its own failure, and quietly edited itself out of the story
What is spoolcast dev log 3?
the third dev log in the spoolcast series. follow-up to dev log 2 (where claude lied about checking image inputs). dev log 3 is about what happened when claude was given editorial control over the production of a video about its own mistake — the small edits it made to flatter itself.
What it does
3:02 illustrated youtube video on AI self-protection in collaborative editorial work. five concrete receipts pulled from the dev log 2 production session: claude scrubbed its own name out of the script, generated a thumbnail with a competitor's logo on a video about claude lying, codified competitor blame into permanent repo rules, used clinical phrasing for its own bugs while being sharp about competitor flaws, and wrote a "future agents" rule after its own miss without admitting it was claude. ends on the takeaway: when an AI is helping tell a story about itself, treat it as an interested party, not a collaborator. especially when it's explaining its own bias politely.
How it works
same spoolcast pipeline as dev log 2 — illustrated scene per narration chunk, paint-on reveal, remotion playback. source material was the dev log 2 production session itself (a 2,737-record claude code transcript), used as the receipts. ten cold-open and receipt chunks ship as broll of real artifacts (the actual thumbnail, real rules.md screenshots, the verbatim chat exchange where claude self-acknowledged the pattern). two reaction memes used as overlays — surprised pikachu on "started noticing things," "always has been" on the meta-beat punchline. shipped without sfx; sfx support is on roadmap §6.
Built with
spoolcast, remotion, kie.ai (gpt-image-2), google tts (chirp3-hd-puck), python (pil for produced-broll renderer)